Developer solves AI hallucination in bank statement tool with balance validation

A developer building StatementDecoder, an AI-powered bank statement converter, found that the hardest challenge was not parsing statements but detecting when the AI output was wrong. Early versions using LLMs to extract structured JSON worked well on sample files but failed on real-world statements from banks like Wise and Revolut, which had unexpected formats and quirks. The developer discovered that prompting improvements and model changes did not solve the core problem of using the same AI to both generate and verify its own answers. The breakthrough came from implementing an independent mathematical validation layer that checks whether opening balances, transactions, and closing balances reconcile — a check that applies regardless of how the data was extracted. This approach meant the system could flag or repair errors before users exported data, shifting the design philosophy toward verifiable AI output rather than perfect AI output.